Soil moisture is a key driver of agricultural productivity. It directly influences plant growth and yield through provision of nutrients to plants. The impact of water availability on global food production is seen as a key global risk and challenge, especially given its over and under supply in many areas. In irrigated production systems it is therefore essential to enhance the efficiency of water use to prevent oversupply and wastage. This requires an understanding of local within-field variations in soil moisture, but techniques to accurately and reliably achieve this are limited.
One potential solution to this problem is to use spaceborne radar systems. Radar backscatter is correlated with soil moisture via the soil’s microwave dielectric constant. While many previous studies have explored these relationships at local or field scale, less effort has been placed on looking at the potential of measuring within-field variations in soil moisture using synthetic aperture radar (SAR). This has been due to the limiting influence of radar speckle as well as lack of long term time series in the past. With the availability of the Copernicus Sentinel-1 system, there now exists the potential to explore these issues further.
The work presented here, explores the capability of Sentinel-1 to monitor sub-field scale soil moisture changes in response to rainfall and irrigation. It also addresses the issue of appropriate sub-field scale ground data for validation, by comparing results from Time Domain Transmissivity (TDT) soil moisture probes and COSMOS cosmic-ray neutron probes (designed by Hydroinova, US). The latter have the advantage that they estimate soil moisture from a radius around a sensor, rather than just at a single point, as is the case with TDT probes. Given that a single point measurement may not be the most appropriate scale for comparison with SAR data, and that factors such as rock and air pockets in the soil can give rise to misleading TDT readings, an understanding of the additional potential offered by the COSMOS system for the validation of SAR data is important.
Using test sites in the UK and Greece, the change in radar backscatter and coherence were investigated and compared to in-situ sensors for fields containing wheat stubble. Shielded COSMOS sensors were mounted on an agricultural robot which recorded soil moisture over a predefined grid of points on dates corresponding with Sentinel-1 overpasses. TDT measurements using 3.8cm and 7.6cm long probes were recorded simultaneously at the same points. At the Greek site, irrigation was applied to manually alter the soil moisture of part of the field and the measurements were repeated. At the UK site, measurements were made in very dry conditions and then repeated following a significant rain event. The minimum area from which reliable soil moisture information could be extracted was investigated by applying a series of different sized windows around each test point. Spatial and temporal variations in both backscatter and coherence were then explored for different overpasses and incidence angles and then compared with results from the IEM model. Results from the experiments are presented and discussed.

Traditional satellite validation requires in situ soil moisture validation networks at scales around 25 km by 25 km. Current and future active microwave missions are pushing this scale down to 100 m or less. Landscape heterogeneity decreases as the resolution increases, especially for agricultural areas with large field sizes. Other land management practices have a spectrum of influences across these scales, so a one size fits all approach to landscape scaling functions may no longer be valid. This study analyzes the ability of a permanent network to estimate field scale soil moisture, using physical sampling and temporary networks. The field site used for this study was the South Fork watershed, which was the study domain of the Soil Moisture Active Passive Validation Experiment 2016 (SMAPVEX16). This is a primarily row crop agriculture watershed with 59 % corn and 24% soybean fields as of 2016. There are twenty semi-permanent soil moisture stations deployed across a pixel of approximately 36 km by 36 km. These sites have been validated at resolutions of 36, 9, and 3 km with various degrees of success. During the summer of 2016, a large scale temporary network was also deployed in the region with several scaling resolutions to provide a greater number of stations with which to compute large scale soil moisture estimates. Field sampling was conducted by scientists during intensive observation periods in the spring and summer of 2016, providing a detailed map of soil moisture variation from the 100 m to 3 km scale. In coordination with the field sampling, aircraft flights were also conducted to collect high resolution maps of soil moisture with less than a 1 km resolution. Results of this study will provide an estimate of the average and variation of soil moisture across these scales with implications for satellite validation in agricultural regions.

Currently, within the Copernicus Global Land Service (https://land.copernicus.eu/global/products), daily Soil Water Index (SWI) information is distributed as gridded product with a resolution of approximately 25 km (Wagner et al., 1999). A surface soil moisture (SSM) product at a resolution equal or lower than 1 km is also under development. Nevertheless, a number of applications, e.g. agriculture, require SSM information at resolutions that correspond to the scale at which water is being used, e.g., “fields scale”. For instance, in irrigated areas, field-to-field variability in SSM, also during early stages of the growing season when fields are still under conditions of low vegetation, is crucial information for improving water management.
For this reason, in the context of the SENSAGRI H2020 project [SENSAGRI.eu], an evolution of the Sentinel-1 (S1) SSM product at 1 km, developed in the contest of the ESA SEOM Exploit-S-1 study [http://seom.esa.int/page_project034.php], has been implemented. The ESA product exploits a change detection approach that requires a short revisit time of SAR observations [Balenzano et al., 2013] and, for this reason, it is referred to as a short-term change detection (STCD) approach. The strength of the algorithm is its conceptual simplicity and robustness. On the other hand, STCD is prone to the occurrence of abrupt changes of the vegetation and/or soil roughness status that can be wrongly interpreted as SSM changes. Such changes, may have a limited impact at a resolution equal or above ~1 km, but they usually produce significant errors at “field scale” resolutions.
The evolution of the ESA product consists of integrating S1 and Sentinel-2 (S2) data in order to improve to “field scale” the resolution of the SSM product. More precisely, S2 NDVI together with S1 is used to mask abrupt changes of the vegetation and/or soil roughness status that may affect SSM estimates at “field scale”. If NDVI is not available (e.g. cloud cover), then the ratio of VH/VV is adopted as a proxy.
The SENSAGRI product transforms dense time series of N co-registered S-1 VV & VH & S-2 NDVI images at 40m pixel size into N-SSM maps. The final step in the SSM retrieval is to apply a low pass filter in order to improve the radiometric accuracy of the final product. In the SENSAGRI product, an adaptive strategy is implemented so that the final product is characterized by a fairly high spatial resolution (~100m) over agricultural areas, where SSM averages and standard deviations at field scale are performed using the European LPIS (Land Parcel Information System) data base for parcel borders, wherever available. Over the areas where parcel borders are not available SSM is averaged at a resolution of 1 km and no S2 data are required.
The paper presents examples of the developed S1 & S2 SSM product over various sites in Europe, reports on the results of the validation activity and provides an outlook about possible applications at regional scale.

Wagner, W., Lemoine, G. & Rott, H. (1999). A Method for Estimating Soil Moisture from ERS Scatterometer and Soil Data. Remote Sensing of Environment. vol.70 (2). pp.191–207.

Balenzano, A., Satalino, G., Lovergine, F., Rinaldi, M., Iacobellis, V., Mastronardi, N. & Mattia, F. (2013). On the use of temporal series of L-and X-band SAR data for soil moisture retrieval. Capitanata plain case study. European Journal of Remote Sensing. vol.46 (1). pp.721–737.

Irrigated agriculture is essential for the global food yield. In the past 40 years, global agricultural production has more than doubled, while the cropland has increased by only 12%, which shows that irrigation has made a great contribution [1,2]. In semi-arid regions, water availability is a major limitation on crop production. Efficient agricultural water management is therefore a major issue, especially in irrigated areas. The development of tools that provide regional estimates of the water balance may help the sustainable management of water resources in such regions [3,4,5]. In this context, research activities show that the synergetic use of radar and optical data have displayed a great potential for the estimation of operational soil surface moisture, directly linked to irrigation dynamic. In this study, we propose the presentation of two irrigation approaches, a first one based on radar Sentinel-1 signal time series and a second one based on soil moisture products retrieved from Sentinel-1 data.
The first approach is applied over Urgell site in Spain (Gao et al., 2018). The backscatter mean value, the signal variance, and the correlation length are derived from the backscatter signal time series and are used for classification. The classification result of irrigated and non-irrigated fields is compared with the ground truth from SIGPAC database over 26,434 fields in the whole study area. The result shows a good overall accuracy, which is 81.08% using the Support Vector Machine (SVM). The mean value of the backscatter time series is the key to separating irrigated and non-irrigated fields, while the correlation length and variance are used for the separation of irrigated trees and irrigated crops. The results of the SVM show good accuracy for irrigated croplands and non-irrigated fields, with soil moisture change more dominant to multi-temporal radar signals (Gao et al., 2017).
The second approach is applied over Kairouan plain site (Bousbih et al., 2018). First, operational approaches were proposed to retrieve soil moisture maps at field scale over the semi-arid Kairouan plain, in the central of Tunisia. The proposed approach is based on the inversion of the Water Cloud Model (WCM) using Sentinel-1 radar data in the VV polarization and NDVI data derived from Sentinel-2 optical images. Retrieved estimations are validated using ground measurements and a Neural Networks (NN) moisture product (El Hajj et al., 2017). Consistent results were found in comparing the two satellite moisture products, with an RMSE of 4.07 vol. % and a bias of 0.56 vol. % between the two estimates. A classification algorithm is proposed for the mapping of irrigated and rainfed areas over the Kairouan plain. The classification first used SVM to distinguish between irrigated and rainfed areas using soil moisture parameters (mean and variance). Next, a Decision Tree classification was used to produce irrigation maps using several simulations based on these parameters for three cases: a classification based on NDVI mean values on agricultural fields; a classification based only on the mean and variance of soil moisture time series; and lastly a classification based on the mean and variance of soil moisture time series, plus NDVI mean values. The classification analysis reveals that the highest overall classification accuracy (77%) was achieved by the classification using soil moisture parameters. The lowest accuracy was retrieved when using only the NDVI mean values.

References
S. Bousbih, M. Zribi, Mohammad El-Hajj, Nicolas Baghdadi, Zohra Lili-Chabaane, Qi Gao and Pascal Fanise, Soil moisture and irrigation mapping in a semi-arid region based on Sentinel-1 and Sentinel-2 synergy, Remote Sensing, 2018.
M. El Hajj, N. Baghdadi, M. Zribi, H. Bazzi, Synergetic use of Sentinel1 and Sentinel2 images for operational soil moisture mapping at high spatial resolution over agricultural areas, Remote Sensing, 9, 1992, 2017.
Q. Gao, M. Zribi, N. Baghdadi, M.J. Escorihuela, Synergetic Use of Sentinel-1 and Sentinel-2 Data for Soil Moisture Mapping at 100 m Resolution, Sensors, 17, 1966; doi:10.3390/s17091966, 2017.
Q. Gao, M. Zribi, M.J. Escorihuela, N. Baghdadi, P. Quintana Segui, Irrigation Mapping Using Sentinel-1 Time Series at Field Scale. Remote Sensing, 10(9), 1495; doi: 10.3390/rs10091495, 2018.

SAR interferometric closure phases signal constitute a kind of inconsistency in triplets of interferograms. Closure phases have been shown to carry information about the propagation in semi-transparent media, typically the moisture status in soils and forests. Simple physical models can explain the phenomenon. Recently, we have developed an inversion algorithm based on closure phases to retrieve the moisture variations from time series of interferometric SAR images. Advantages of this method are high resolution, compared to existing methods, and requirement of only single polarization data [1].
Low noise interferograms are necessary when inverting closure phases to estimate soil or vegetation moisture. A stack of highly coherent L-band images with short revisit time would be ideal. Short spatial baselines are also required to reduce fine topographic effects on the interferometric phase. Unfortunately, the availability of these data is rather limited. Only few ALOS-2 image series with the desired characteristics are available, as the ALOS-2 acquisition plan privileges global coverage instead of building stacks on specific points: this implies long temporal baselines between images. ALOS stacks present the same issue with the addition of worse spatial baselines. To better distinguish the phase response of different target types with respect to moisture changes, high resolution is also a desirable feature.
The UAVSAR is an airborne L-band synthetic aperture radar operated by JPL. UAVSAR data are precisely calibrated and can be used for interferometric applications; small baselines are ensured by an autopilot system. There are several interferometric UAVSAR stacks with short revisit time. High resolution (range and azimuth respectively 1.8 and 0.8 meters) simplifies the separation and understanding of different ground coverage and shows the potential of the method.
We will present experimental results with UAVSAR data: examples of closure phases for different soil types and ground coverages and examples of moisture estimation based on closure phases. Crossing flight lines acquired during same days over the Sacramento valley show the consistency of phase closure with respect to a change of geometry. Since data are quad-polarized, we analyze polarization-dependent distortions of the closure phase to establish their effects on moisture estimation. UAVSAR data has been used for the Can-Ex10 and SMAPVEX12 validation campaigns of the Soil Moisture Active Passive (SMAP) satellite, which provides measurements of the land surface soil moisture. Measurements with moisture probes have been recorded during those campaigns and can be used for local validations. UAVSAR stacks can also be used with current methods that exploit the backscatter of images with different polarization to estimate soil moisture. We will thus compare our inversion results with moisture probes data, backscatter-based moisture estimates from the same data, low-resolution satellite-based remote sensing products, and with large scale models.
A first example of our experiments can be found in the attached image. Figure 1 shows an example of phase closure for a six images UAVSAR dataset acquired in the San Joaquin valley. Figure 2 is the estimated moisture for one date, inverted from the San Joaquin dataset. Figure 3 is a comparison of moisture estimates for a shrubland field in the San Joaquin valley; this plot is adapted from [2]. The moisture estimation in [2] is based on the backscatter of the SAR images (red line). The superimposed green dots are our moisture estimates obtained from closure phases.

[1] De Zan, Francesco, and Giorgio Gomba. "Vegetation and soil moisture inversion from SAR closure phases: First experiments and results." Remote Sensing of Environment 217 (2018): 562-572.
[2] Kim, S.B., M. Arii, T. Jackson, Modeling L-band synthetic aperture radar data through dielectric changes in soil moisture and vegetation over shrublands, IEEE J. Selected Topics Applied Earth Obs. Remote Sensing, DOI: 10.1109/JSTARS.2017.2741497, 2017

Due to its relation to the Earth’s climate and weather, as well as to phenomena like drought, flooding, or land-slides, knowledge of the soil moisture content is valuable to many scientific and professional (e.g. non-scientific domain experts or policy makers) users. Satellite remote sensing presents the only possibility for the spatially continuous measurement of the soil moisture content over large areas. Estimations using microwave remote sensing systems have evolved over the last few decades and nowadays several operational, coarse resolution products are available, relying on passive or active microwave sensors like SMOS (passive), SMAP (passive), or ASCAT (active). With the openly available imagery of the Copernicus C-Band SAR system Sentinel-1 (S1) – offering high spatial (20 m), high temporal resolution (6-day revisit) data with a near-global coverage – the technical foundation for the high-resolution mapping of surface soil moisture was created. With this contribution we want to address some of the challenges we are facing in pursuit of this goal: How to solve the retrieval problem? How can we manage the increasingly large amounts of satellite data? What is the optimal resolution?
To address the first point, we introduce a purely data-driven, Support Vector Regression (SVR) based approach. Machine learning techniques allows building a retrieval model based on the relationship between in-situ measurement, S1 intensity and auxiliary data. The main challenge for this task was to find an appropriate training data-set and design a model training scheme to derive a universal retrieval model for the global application. Common problems like errors due to extrapolation, site-specificity of the model, or the loss of target variability were to be avoided. The International Soil Moisture Network (ISMN) is an international cooperation to establish and maintain a global in-situ soil moisture database. An online platform bundles the data from hundreds of monitoring stations worldwide from many network providers. Data from the ISMN was used as a reference data-set for the SVR training. It was found to serve the purposes well, giving a good representation of different climate zones. The most important novelty of the presented method is a two-step training and estimation approach. Average conditions and relative soil moisture are estimated separately, which allows to correctly represent spatial and temporal variability.
Over land, S1 A and B produce more than 1 TB of data daily, which makes it increasingly difficult for the user to perform post-processing. Several providers of pre-processed Satellite data try to offer a solution for this problem. One of these providers is Google with its Google Earth Engine (GEE). For our study we have used the pre-processed IW GRDH product. The following processing steps are applied by Google:
1. Application of precise orbit file
2. Thermal noise removal
3. Radiometric calibration
4. Terrain correction using SRTM 30 or ASTER DEM for areas greater than +/- 60° latitude, where SRTM is not available
For the retrieval of the soil moisture content we implemented the following further processing steps, which are applied on-the-fly, on retrieval of the data via GEE.
1. Static masking of extreme backscatter and local incidence angle values
2. Dynamic masking of snow and frozen soil
3. Multi-temporal speckle filtering
The GEE S1 product, used in this study is sampled to a global 10 m grid, which would allow very high-resolution soil moisture content estimations. But what is the thematic resolution of the data? To address this question, the data were resampled to different spatial resolutions (5000 m, 2000 m, 1000 m, 500 m, 100 m, 50 m, 10 m) and used to train 7 dedicated retrieval models.
Globally, the best average accuracy of estimated soil moisture is Root-Mean-Squared-Error of 0.07 m3m-3 and a correlation, R, of 0.84 (computed based a set of ISMN stations, which were not used for the model training). Preliminary results indicate that the overall accuracy does not increase significantly, for spatial resolutions higher than 500 m. However, in the case of individual stations the highest retrieval accuracy was achieved with a spatial resolution of 50 m. At the full resolution of 10 m noise seems to prevent any further improvement. The analysis shows that the “optimal” resolution strongly depends on the land-cover type and vegetation dynamics.
This study is carried out in the framework of two projects: 1) The Horizon 2020 project “Ecopotential – Improving Future Ecosystem Benefits through Earth Observation, which has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement n° 641762); 2) “Synergy of multi-temporal Sentinel 1 and Sentinel 2 image for soil water resources monitoring in Africa”, funded within the European Space Agency “Tiger Bridge – Water for Agriculture” framework.